Crypto & Investing — The 10-Agent Pipeline
The 10-Agent Pipeline¶
Every /invest analyze <ticker> runs the same pipeline: 10 specialized Claude subagents across 4 phases. The design is adapted from TauricResearch/TradingAgents but rebuilt around Claude Code primitives (subagents, Memory MCP, MCP tool use).
graph TD
A["/invest analyze VTI"] --> B["Phase 1: Data Gathering (parallel)"]
B --> B1["1. Market Analyst"]
B --> B2["2. News Analyst"]
B --> B3["3. Sentiment Analyst"]
B --> B4["4. Fundamentals Analyst"]
B1 & B2 & B3 & B4 --> C["Phase 2: Investment Debate (sequential)"]
C --> C1["5. Bull Researcher"]
C --> C2["6. Bear Researcher"]
C1 & C2 --> C3["7. Research Manager"]
C3 --> D["Phase 3: Trade Proposal"]
D --> D1["8. Trader"]
D1 --> E["Phase 4: Risk Assessment"]
E --> E1["9. Risk Panel (3-way debate)"]
E1 --> E2["10. Portfolio Manager"]
E2 --> F["Educational Layer"]
F --> G["Discord + Memory MCP"]
Phase 1 — Data Gathering (Parallel)¶
Four analysts run simultaneously, each with their own tool access. Parallelism cuts wall-clock time roughly 4×.
| # | Agent | Role | Data Sources |
|---|---|---|---|
| 1 | Market Analyst | OHLCV prices, technical indicators (SMA, EMA, MACD, RSI, Bollinger, ATR, VWMA, MFI) | Alpha Vantage, Finnhub, Finviz via crawl4ai |
| 2 | News Analyst | Breaking news, macro events, earnings surprises, insider transactions | WebSearch, crawl4ai, Playwright |
| 3 | Sentiment Analyst | Social buzz, retail attention, hype detection | YouTube Data API v3, Google Trends, Reddit (via WebSearch + crawl4ai), Finnhub social |
| 4 | Fundamentals Analyst | P/E, revenue, margins, balance sheet, cash flow, dividend history | Alpha Vantage, Finnhub, Yahoo Finance via crawl4ai |
Phase 2 — Investment Debate (Sequential)¶
The four analyst reports feed a three-way debate.
| # | Agent | Role | Input |
|---|---|---|---|
| 5 | Bull Researcher | Build the strongest possible case FOR investing | All 4 analyst reports + Memory MCP past reflections |
| 6 | Bear Researcher | Build the strongest possible case AGAINST investing | All 4 analyst reports + Memory MCP past reflections |
| 7 | Research Manager | Judge the debate and pick a direction with explicit reasoning | Bull + Bear arguments |
Why separate bull and bear agents?
A single agent asked "is this a good buy?" tends to rationalize either direction. Forcing separate agents to each build the strongest case for their side, then a third to judge, consistently surfaces weaknesses a single analyst would miss.
Phase 3 — Trade Proposal¶
| # | Agent | Role | Input |
|---|---|---|---|
| 8 | Trader | Concrete entry price, exit targets, position sizing | Research Manager's thesis + portfolio context |
The Trader turns the thesis into actionable numbers: "30% tranche at market, scale the remaining 70% over 60-90 days via DCA, stop-loss at -12%."
Phase 4 — Risk Assessment (3-Way Debate)¶
| # | Agent | Role | Input |
|---|---|---|---|
| 9 | Risk Panel | Three perspectives — Aggressive, Conservative, Neutral — argue the Trader's proposal | Trader proposal + all prior reports |
| 10 | Portfolio Manager | Final Buy / Overweight / Hold / Underweight / Sell decision | Everything above + your Ghostfolio holdings |
The Portfolio Manager is the only agent that sees your actual portfolio. Its rating factors in concentration risk, sector exposure, and available cash from Firefly III.
Post-Pipeline — Educational Layer¶
Not a separate agent — a post-processing pass over the Portfolio Manager's output that:
- Translates thesis into beginner-friendly language (calibrated to your Foundation/Growth/Advanced stage)
- Explains every metric and indicator used (P/E, RSI, VWMA, etc.)
- Adds a "Concept of the Week" pulled from what showed up in this analysis
- Adds "What to watch for" next steps (earnings dates, Fed meetings, sector events)
Output Routing¶
| Destination | What Goes There |
|---|---|
| Memory MCP | Always — primary data store for past analyses and reflections |
| Discord (Apprise) | Always — real-time alert with rating, thesis, concept |
| Obsidian | Only weekly/monthly reports, learning notes, reflections (analyses go stale within hours) |
| Monthly report only |
TL;DR
10 agents across 4 phases: 4 analysts gather data in parallel, Bull and Bear debate, Research Manager judges, Trader proposes entry/exit/sizing, Risk Panel runs a 3-way debate, Portfolio Manager renders a final rating with your actual holdings in mind. Educational layer post-processes everything for your learning stage. Memory MCP + Discord always, Obsidian/PDF only for summaries.
Previous: ← Getting Started | Next: Finance Stack Integration →